wizz-method 2.17.1 → 2.18.0

package/src/skills-lib/wizz-video-styles/SKILL.md

@@ -20,6 +20,8 @@ It can also run at step 1, so the script is written for the style (room for b-ro
 
 ## Per piece
 
+Scripts live in this skill's `scripts/`; run them from the project root (the history file is written in the current directory, or in `--project <dir>`).
+
 1. **Inventory first, always.** Ask or read from the plan, one line each:
    - Is a real person on camera? Which shots exist (face to camera, podcast, stage, street, hands only)?
    - Voice: the person's own recording, a podcast/interview track, TTS, or none?
@@ -27,16 +29,17 @@ It can also run at step 1, so the script is written for the style (room for b-ro
    - B-roll or stock, a physical product, food, a place?
    - Niche, goal of the piece (follow, comment, save, click), duration.
    Never pick a style the material cannot feed. "Talking head" needs a person on camera; "motion editorial" needs nothing but a voice; "macro / ASMR" needs a physical product filmed close.
-2. **Pick the style** with [references/style-picker.md](references/style-picker.md): material decides which styles are possible, niche and goal rank them. Mixed is allowed and common: person on camera for the hook and the CTA, another style in the middle.
+2. **Pick the style** with [references/style-picker.md](references/style-picker.md): material decides which styles are possible, niche and goal rank them. Mixed is allowed and common: person on camera for the hook and the CTA, another style in the middle. Before deciding, run `node scripts/references.mjs history` in the project: if the last two pieces used the same style, do not pick it a third time unless the user wants a series; choose another style the material allows, or keep it only for the hook and CTA.
 3. **Open the style card** in `references/styles/` and follow it: rhythm, captions, palette, typography, motion, music, SFX, hook, CTA, and its "do not" list.
-4. **Watch two of its references** (links in the card) before editing. Reading a card is not watching a video.
+4. **Get fresh references and watch two** before editing: `node scripts/references.mjs pick --style <key>` returns measured references plus, for styles generated in code, current viral videos from the awesome-opus5-5-videos list, skipping everything this project already used. Reading a card is not watching a video. For how those generated videos were briefed, see [references/generated-prompts.md](references/generated-prompts.md).
 5. **Audio:** music and SFX choices from [references/audio.md](references/audio.md); the redistributable kit is in `assets/audio/` (credits in `assets/audio/CREDITS.md`).
 6. **Hand over** to `wizz-video-effects` with the card's SFX ceiling and music level, then to Remotion.
 7. **Look at the final file**: extract frames and compare them side by side with a contact sheet of one reference.
+8. **Record the piece**: `node scripts/references.mjs record --style <key> --piece <slug> --refs <ids watched>`. It writes `_wizz-output/video-styles/historico.json`, which is what keeps the next piece from repeating the same references and look.
 
 ## What the measurements say across every style
 
-- **SFX are rare.** Across 99 measured shorts, sound effects landed on a cut in a small minority of cuts (per-style counts in each card). Putting a sound on every cut is the most common amateur tell.
+- **SFX are rare.** Across 104 measured shorts, sound effects landed on a cut in a small minority of cuts (per-style counts in each card). Putting a sound on every cut is the most common amateur tell.
 - **Music bed: style-dependent.** Fast montage, animated explainers and travel POV carry a continuous bed; talking heads and podcast clips mostly have none or a faint one.
 - **Captions** in reach-heavy styles are 1 to 3 words at a time, bold, lower-centre, with one keyword in an accent colour; editorial and explainer styles keep them small or drop them.
 - **Rhythm is a style property, not a virtue.** The highest-reach videos in the set range from single takes with no detected cut to a cut every second.
@@ -53,7 +56,10 @@ Details and the per-style table: [references/measured-patterns.md](references/me
 | [references/audio.md](references/audio.md) | music and SFX per style, the kit, libraries and licences |
 | [references/research-method.md](references/research-method.md) | how the catalogue was measured, its limits, how to add a niche or a style |
 | `data/references.json` | every measured short: link, channel, views, niche, style, metrics |
+| [references/generated-prompts.md](references/generated-prompts.md) | prompts of viral code-generated videos, grouped by style, and how they run here |
 | `scripts/analyze.sh` | measures one short (cuts, music bed, SFX heuristics, loudness, contact sheet) |
+| `scripts/references.mjs` | fresh references per style without repeats, and the project's history of styles used |
+| `scripts/beat_grid.py` | BPM, beats and downbeats of a song, so cuts snap to the music (lyric video, kinetic typography) |
 | `assets/remotion/editorial/` | Remotion template of the motion editorial style |
 | `assets/audio/` | licensed music beds and SFX that may ship with a project |
 

package/src/skills-lib/wizz-video-styles/data/recipes.json

@@ -7,7 +7,8 @@
   "e": "health / fitness / nutrition",
   "f": "food / recipes / lifestyle / travel",
   "g": "beauty / fashion / local services (clinic, salon, architecture, real estate)",
-  "h": "education / science / comedy"
+  "h": "education / science / comedy",
+  "i": "music / artists / lyric videos"
  },
  "styles": {
   "motion-editorial": {
@@ -336,6 +337,37 @@
    "dont": [
     "Setup longer than 3 s before the first joke."
    ]
+  },
+  "tipografia-cinetica": {
+   "title": "Kinetic typography and lyric video (mexicat P(doom), PixelValet, Ninh Truong)",
+   "needs": "An audio track with words: a song with its lyrics, or a voice-over with its script. Nobody on camera, no footage. Everything is generated (HyperFrames, three.js, canvas). Word timing from ctc-align; a beat grid when there is music.",
+   "fits": "Music releases and lyric videos, manifestos, a strong opinion read aloud, brand promos built on a slogan, poetry. Shares and rewatches; weak for teaching step by step.",
+   "look": [
+    "The words ARE the image: one keyword huge in the frame, the rest of the line small. No caption strip at the bottom.",
+    "Plain background (black, or a flat colour that changes per section) or one generated world the camera travels through (a map, a grid in perspective, a 3D object the line wraps around).",
+    "Each word is animated by what it means: PAIN breaks, CHEERFUL bounces, LIGHT lights up, a word curves around the object it names.",
+    "Code-rendered pieces add film treatment: bloom, halation, grain, a thin HUD in mono."
+   ],
+   "palette": "Dark base (#0A0A0A or a deep teal #0F3B3A) plus ONE glowing accent (orange #FF6A1A in P(doom), coral in Maps) and off-white text. Flat-colour variant: one saturated background per word or section, text in white or ink.",
+   "type": "Two or three families with high contrast: a heavy grotesk or condensed caps for the keyword (Archivo Black, Bebas Neue), a mono for small lines and HUD (IBM Plex Mono), optionally a serif italic for a soft word (Cormorant Garamond). Mix weights inside the same line.",
+   "motion": [
+    "Words enter on the syllable they are sung or spoken (forced alignment, never guessed); the keyword lands on the beat.",
+    "Scene changes snap to the downbeat. P(doom) changes every ~0.3 s at peaks; the calm variant is a single continuous camera move for a whole verse.",
+    "The camera moves through the type (push, orbit, travel across a map) instead of the text sliding over a still frame.",
+    "Hold the last word of a line a beat longer before it leaves."
+   ],
+   "captions": "None separate: the type is the caption. Keep the full line readable somewhere (small, mono, corner) when the keyword alone would lose the sense.",
+   "hook": "The first sung or spoken word, huge, in the accent colour, landing on the first downbeat (0 to 1 s). No logo intro.",
+   "cta": "The title or slogan locked up big and still for 1 to 2 s, the accent lit; one small line (artist, handle, link) under it.",
+   "music": "The track itself is the whole soundtrack. Lyric videos: song at full level, no extra bed. Voice-over promos: a beat-driven bed under the voice.",
+   "sfx": "Song-driven pieces: almost none (5 of 56 cuts in P(doom), 0 of 2 in Maps). Promo variant: hits on every word change (9/13 and 16/20 cuts), which the heuristic partly confuses with the kick. Never add whooshes over a song.",
+   "dont": [
+    "Karaoke subtitles at the bottom over a static background: that is a lyric 'overlay', not this style.",
+    "Guessing word timing by ear; align it (ctc-align) and snap scenes to the beat grid.",
+    "More than one accent colour, or a different font for every word.",
+    "Shipping 16:9: every measured reference is horizontal; recompose for 9:16 with the keyword in the upper-middle and nothing in the bottom 20%."
+   ],
+   "template": null
   }
  }
 }

package/src/skills-lib/wizz-video-styles/data/references.json

@@ -3393,5 +3393,180 @@
    "on_cuts": "0/2"
   },
   "lufs": "-14.4"
+ },
+ {
+  "id": "5EoO5413dBY",
+  "url": "https://www.youtube.com/watch?v=5EoO5413dBY",
+  "channel": "mexicat",
+  "views": "120706",
+  "nicho": "i",
+  "idioma": "en",
+  "style": "tipografia-cinetica",
+  "estilo_visual": "Clipe musical renderizado em código (three.js): fundo preto, uma palavra ou verso gigante por batida sobre cenas 3D (buraco negro, criatura, torre, grade em perspectiva), HUD de interface e formulários de papel",
+  "motion_transicoes": "Troca de cena no downbeat, texto entra palavra a palavra no tempo da voz, câmera orbita objetos 3D, bloom e grão de filme, texto curvado em volta de objetos",
+  "legenda": "A letra É o visual: palavra-chave gigante (sans bold ou serifa) + verso pequeno em mono no canto; nunca legenda embaixo",
+  "hook_0_3s": "Linhas de código de terminal e uma palavra laranja enorme ('AGI') já no primeiro verso",
+  "paleta": "preto, laranja incandescente, branco/creme; um respiro em papel bege",
+  "tipografia": "sans grotesca pesada (Archivo), mono (IBM Plex Mono), serifa itálica (Cormorant) e fonte de traço único",
+  "ritmo": {
+   "cuts_per_10s": 3.57,
+   "shot_median_s": 0.27
+  },
+  "musica": {
+   "guess": "continua",
+   "ratio": 0.326,
+   "impressão": "heurística de áudio; nome da faixa não confirmado"
+  },
+  "sfx": {
+   "counts": {
+    "pop_click": 0,
+    "whoosh": 29,
+    "impact_low": 0
+   },
+   "per_10s": 1.85,
+   "on_cuts": "5/56"
+  },
+  "lufs": "-15.5",
+  "formato": "16:9 (horizontal); adaptar para 9:16"
+ },
+ {
+  "id": "FOMWS5pChuA",
+  "url": "https://www.youtube.com/watch?v=FOMWS5pChuA",
+  "channel": "Ninh Truong",
+  "views": "243944",
+  "nicho": "i",
+  "idioma": "en",
+  "style": "tipografia-cinetica",
+  "estilo_visual": "Lyric video em tipografia cinética: palavras da letra montadas como objetos num mundo 2D (mapa de ruas visto de cima, nuvens, rádio), fundo verde-petróleo",
+  "motion_transicoes": "Câmera viaja e gira pelo mapa seguindo o verso, palavras dobram em 3D e se encaixam nas ruas, poucas trocas de cena (plano contínuo)",
+  "legenda": "A letra é o visual: verso inteiro, palavras de pesos e tamanhos misturados, palavra-chave em destaque de cor",
+  "hook_0_3s": "Letra 'I' gigante no escuro, depois o verso se monta em blocos",
+  "paleta": "verde-petróleo escuro, verde-menta, coral/salmão de acento, cinza nas cenas de céu",
+  "tipografia": "sans condensada em caixa alta misturada com sans fina e script; pesos alternados no mesmo verso",
+  "ritmo": {
+   "cuts_per_10s": 0.63,
+   "shot_median_s": 9.92
+  },
+  "musica": {
+   "guess": "continua",
+   "ratio": 0.218,
+   "impressão": "heurística de áudio; nome da faixa não confirmado"
+  },
+  "sfx": {
+   "counts": {
+    "pop_click": 0,
+    "whoosh": 2,
+    "impact_low": 0
+   },
+   "per_10s": 0.63,
+   "on_cuts": "0/2"
+  },
+  "lufs": "-15.9",
+  "formato": "16:9 (horizontal); adaptar para 9:16"
+ },
+ {
+  "id": "Kt8vGm8_ekQ",
+  "url": "https://www.youtube.com/watch?v=Kt8vGm8_ekQ",
+  "channel": "PixelValet",
+  "views": "928526",
+  "nicho": "h",
+  "idioma": "en",
+  "style": "tipografia-cinetica",
+  "estilo_visual": "Uma palavra por cena (emoções) no centro de um fundo de cor chapada; cada palavra é animada pelo significado dela (PAIN quebra, CHEERFUL pula em arco, LIGHT acende)",
+  "motion_transicoes": "Uma bolinha guia passa de cena em cena; troca de fundo a cada palavra; a letra se deforma conforme o sentido",
+  "legenda": "Sem legenda; a palavra é o único elemento",
+  "hook_0_3s": "Fundo roxo, bolinha e a primeira palavra se desenhando",
+  "paleta": "cores chapadas que mudam por emoção (roxo, amarelo, azul, verde, vermelho, cinza escuro)",
+  "tipografia": "sans geométrica, uma família só, variando peso e espaçamento",
+  "ritmo": {
+   "cuts_per_10s": 0.68,
+   "shot_median_s": 7.29
+  },
+  "musica": {
+   "guess": "sem",
+   "ratio": 0.021,
+   "impressão": "heurística de áudio; nome da faixa não confirmado"
+  },
+  "sfx": {
+   "counts": {
+    "pop_click": 0,
+    "whoosh": 78,
+    "impact_low": 16
+   },
+   "per_10s": 21.18,
+   "on_cuts": "1/3"
+  },
+  "lufs": "-22.5",
+  "formato": "16:9 (horizontal); adaptar para 9:16"
+ },
+ {
+  "id": "9ngNySFSWWU",
+  "url": "https://www.youtube.com/watch?v=9ngNySFSWWU",
+  "channel": "PiXeL ARTWORKS",
+  "views": "1215529",
+  "nicho": "d",
+  "idioma": "en",
+  "style": "tipografia-cinetica",
+  "estilo_visual": "Texto pequeno branco no centro de fundo preto, alternando com fundo branco de papel; nada além do texto",
+  "motion_transicoes": "Corte seco a cada frase no ritmo da narração/batida; palavras trocam de lugar e de caixa; sem objetos",
+  "legenda": "O texto é a narração escrita, uma frase curta por vez",
+  "hook_0_3s": "Logo do estúdio, depois 'fine good? ok' no preto",
+  "paleta": "preto, branco, um vermelho só na abertura",
+  "tipografia": "sans neutra pequena com tracking largo",
+  "ritmo": {
+   "cuts_per_10s": 4.62,
+   "shot_median_s": 1.54
+  },
+  "musica": {
+   "guess": "sem",
+   "ratio": 0.0,
+   "impressão": "heurística de áudio; nome da faixa não confirmado"
+  },
+  "sfx": {
+   "counts": {
+    "pop_click": 0,
+    "whoosh": 44,
+    "impact_low": 78
+   },
+   "per_10s": 28.2,
+   "on_cuts": "16/20"
+  },
+  "lufs": "-8.9",
+  "formato": "16:9 (horizontal); adaptar para 9:16"
+ },
+ {
+  "id": "A0CKFb24MJ8",
+  "url": "https://www.youtube.com/watch?v=A0CKFb24MJ8",
+  "channel": "Cinthya Studios",
+  "views": "353630",
+  "nicho": "d",
+  "idioma": "en",
+  "style": "tipografia-cinetica",
+  "estilo_visual": "Promo em tipografia cinética: palavras em caixa alta condensada que preenchem a tela, algumas com vídeo de cidade dentro das letras",
+  "motion_transicoes": "Corte na batida, palavras escalam e deslizam, texto vira máscara de vídeo, ícone animado no fim",
+  "legenda": "O texto é a mensagem de venda; uma ou duas palavras por vez",
+  "hook_0_3s": "'CONVERT' em condensada preta no branco, inverte pra fundo preto",
+  "paleta": "preto, branco, azul e vermelho só no logo final",
+  "tipografia": "sans condensada em caixa alta (estilo Bebas), pesos variando",
+  "ritmo": {
+   "cuts_per_10s": 5.54,
+   "shot_median_s": 1.32
+  },
+  "musica": {
+   "guess": "continua",
+   "ratio": 0.176,
+   "impressão": "heurística de áudio; nome da faixa não confirmado"
+  },
+  "sfx": {
+   "counts": {
+    "pop_click": 0,
+    "whoosh": 0,
+    "impact_low": 46
+   },
+   "per_10s": 19.6,
+   "on_cuts": "9/13"
+  },
+  "lufs": "-10.7",
+  "formato": "16:9 (horizontal); adaptar para 9:16"
  }
 ]

package/src/skills-lib/wizz-video-styles/data/styles.json

@@ -722,5 +722,60 @@
     "url": "https://www.youtube.com/shorts/zKJSyVJB_yE"
    }
   ]
+ },
+ "tipografia-cinetica": {
+  "n": 5,
+  "nichos": [
+   "d",
+   "h",
+   "i"
+  ],
+  "langs": [
+   "en"
+  ],
+  "views_max": 1215529,
+  "views_median": 353630,
+  "cuts_per_10s": [
+   0.63,
+   3.57,
+   5.54
+  ],
+  "shot_median_s": 1.54,
+  "music": {
+   "continua": 3,
+   "fraca": 0,
+   "sem": 2
+  },
+  "sfx_on_cuts": [
+   31,
+   94
+  ],
+  "top": [
+   {
+    "channel": "PiXeL ARTWORKS",
+    "views": "1215529",
+    "url": "https://www.youtube.com/watch?v=9ngNySFSWWU"
+   },
+   {
+    "channel": "PixelValet",
+    "views": "928526",
+    "url": "https://www.youtube.com/watch?v=Kt8vGm8_ekQ"
+   },
+   {
+    "channel": "Cinthya Studios",
+    "views": "353630",
+    "url": "https://www.youtube.com/watch?v=A0CKFb24MJ8"
+   },
+   {
+    "channel": "Ninh Truong",
+    "views": "243944",
+    "url": "https://www.youtube.com/watch?v=FOMWS5pChuA"
+   },
+   {
+    "channel": "mexicat",
+    "views": "120706",
+    "url": "https://www.youtube.com/watch?v=5EoO5413dBY"
+   }
+  ]
  }
 }

package/src/skills-lib/wizz-video-styles/references/generated-prompts.md

@@ -0,0 +1,60 @@
+# Generated-video prompts, by catalogue style
+
+Fifteen prompts picked from [awesome-opus5-5-videos](https://github.com/yihui-dev/awesome-opus5-5-videos) (513 entries in its `data/videos.json`, MIT, checked on 2026-10-10): videos that a model wrote as code (canvas, SVG, three.js) and that went viral. Each line says what the prompt asks for, in our words; the full prompt and the video are at the links. The prompts belong to their creators: link to them, quote at most a phrase, never paste the whole prompt into a project or a skill.
+
+This page is a fixed sample. For current videos, and to skip the ones a project already used, run `node scripts/references.mjs pick --style <key>`.
+
+Use this page **after** the style is chosen ([style-picker.md](style-picker.md)), to see how others briefed a model for that look. These are not measured references: none of them enter `data/references.json` unless one is measured with `scripts/analyze.sh`.
+
+## How a generated piece runs in the Method
+
+- **HyperFrames draws, Remotion assembles** (`_shared/video-pipeline.md`, wizz module). A prompt that asks for "one HTML file with the music" becomes: scenes drawn silent in HyperFrames, assembled in Remotion with the audio.
+- **Timing is measured, never guessed.** Words from `ctc-align`; cuts on the downbeats from [`scripts/beat_grid.py`](../scripts/beat_grid.py). Several prompts below ask for a "120 BPM grid" by hand; the script gives the real grid of the actual song.
+- **Recompose for 9:16** when the original was square or horizontal: keyword in the upper-middle, nothing important in the bottom 20 %.
+
+## Kinetic typography and lyric video ([card](styles/tipografia-cinetica.md))
+
+| Prompt | What it asks | Links |
+|---|---|---|
+| anabology · Claude Pop remake | Reimagine a whole pop song's video end to end with the same audio, every lyric represented on screen, free to go abstract. | [post](https://x.com/anabology/status/2103534482930491441) · [video](https://skillry.dev/ai-videos/opus-5-5/anabology-491441) |
+| doubleunplussed · music video "like P(doom)" | Music video in the P(doom) style for another song, with references and lore instead of literal lyrics ("tokens" becomes a stochastic parrot). | [post](https://x.com/doubleunplussed/status/2103697580421181894) · [video](https://skillry.dev/ai-videos/opus-5-5/doubleunplussed-181894) |
+| ParanoidAmerica · evil pizza parlour | A full music video synced to a `lyrics.txt`, one continuous camera journey through a place, collage plus ballpoint-pen look. | [post](https://x.com/ParanoidAmerica/status/2103570879619686717) · [video](https://skillry.dev/ai-videos/opus-5-5/paranoidamerica-686717) |
+| twoclipping · wordmark launch film | A one-word brand built in one continuous take to a 120 BPM song with a drop and a breakdown; every scene grows out of the previous one, nothing fades or cuts. | [post](https://x.com/twoclipping/status/2103835273813496100) · [video](https://skillry.dev/ai-videos/opus-5-5/twoclipping-496100) |
+
+## Motion editorial and product UI ([card](styles/motion-editorial.md))
+
+| Prompt | What it asks | Links |
+|---|---|---|
+| twoclipping · one shape, many UI states | One element morphs through 8 to 12 UI states (button, loader, player, chart), a cursor drives each change, springs, cut on a 120 BPM grid. | [post](https://x.com/twoclipping/status/2103273003555402193) · [video](https://skillry.dev/ai-videos/opus-5-5/twoclipping-402193) |
+| verbove · product story in one canvas | The same morphing-shape idea as strict rules: one `draw(t)` function, real UI and real data, something on every beat, exports in 1:1, 16:9 and 9:16. | [post](https://x.com/verbove/status/2103483957266268381) · [video](https://skillry.dev/ai-videos/opus-5-5/verbove-268381) |
+| AnnaCher · promo from the product's own UI | Same structure built from the real product UI, one accent colour, last frame equals the first so it loops; a long banned list (glows, particles, bouncy easing). | [post](https://x.com/AnnaCher___/status/2103571096549433425) · [video](https://skillry.dev/ai-videos/opus-5-5/annacher-433425) |
+| jhylee95 · launch video | Research the product, hook in the first second, distinct beats for proof, every cut and hit retimed until it lands on a beat or drop. | [post](https://x.com/jhylee95/status/2103604431627452427) · [video](https://skillry.dev/ai-videos/opus-5-5/jhylee95-452427) |
+| twoclipping · minimal launch with real footage | 10 bars at 120 BPM, one idea per shot, masked type reveals and match cuts over the owner's real vertical clips; no placeholder cards. | [post](https://x.com/twoclipping/status/2102554209166000267) · [video](https://skillry.dev/ai-videos/opus-5-5/twoclipping-000267) |
+| Mounnna · logo reveal | Logo drawn as a wireframe, low-poly shards fly in and assemble it, then the icon background and the typed brand name. A 3 to 5 s sting for the CTA. | [post](https://x.com/Mounnna/status/2103802871934497266) · [video](https://skillry.dev/ai-videos/opus-5-5/mounnna-497266) |
+
+## Animated explainer ([card](styles/explicador-animado.md))
+
+| Prompt | What it asks | Links |
+|---|---|---|
+| alex_prompter · 30 s business explainer | Five scenes: the customer's problem, what you do, how it works in 3 steps, one proof point, your name. The simplest brief that works. | [post](https://x.com/alex_prompter/status/2103499977632997524) · [video](https://skillry.dev/ai-videos/opus-5-5/alex-prompter-997524) |
+| emollick · recursion | Explain one idea where each explanation switches to a radically different visual style; self-referential and fast. | [post](https://x.com/emollick/status/2103688362960019567) · [video](https://skillry.dev/ai-videos/opus-5-5/emollick-019567) |
+| koldo2k · infinite zoom loop | A looping camera flight from landscape to landscape through vintage objects, collage and halftone look, film grain. | [post](https://x.com/koldo2k/status/2103129343253778767) · [video](https://skillry.dev/ai-videos/opus-5-5/koldo2k-778767) |
+
+## Whiteboard ([card](styles/quadro-branco.md))
+
+| Prompt | What it asks | Links |
+|---|---|---|
+| tak3sh8 · hand-drawn physics | Hand-drawn whiteboard animation of a hard concept, then a second prompt turning it into an MP4 with narration. | [post](https://x.com/tak3sh8/status/2103667481139441895) · [video](https://skillry.dev/ai-videos/opus-5-5/tak3sh8-441895) |
+
+## Fast meme ([card](styles/meme-rapido.md))
+
+| Prompt | What it asks | Links |
+|---|---|---|
+| kloss_xyz · 9:16 brain rot | A chaotic vertical video generated in Python and rendered with ffmpeg, with motion and sound design, told from the model's point of view. | [post](https://x.com/kloss_xyz/status/2103664956482941143) · [video](https://skillry.dev/ai-videos/opus-5-5/kloss-xyz-941143) |
+
+## What the good ones share
+
+- **Inputs first.** The strongest prompts open with an `<inputs>` block: ask for the real product, the real UI states, the real data, the brand colour and the song before drawing anything.
+- **A banned list.** Glows, particle bursts, RGB split, camera shake, bouncy easing: naming what not to do does more than adjectives.
+- **One rule for continuity.** "One shape, never cut" or "every scene is made out of the previous one" gives the piece its identity.
+- **The beat as the clock.** Every change lands on a beat or a drop. Here that clock comes from `beat_grid.py`, not from a hand-typed BPM.

package/src/skills-lib/wizz-video-styles/references/measured-patterns.md

@@ -2,6 +2,8 @@
 
 Set measured on 2026-10-04: **99 public shorts**, 46 in English and 53 in Brazilian Portuguese, 8 niche groups (9 to 16 per group), views from 23,718 to 107,324,401 (median 608,820). Per-style numbers live in [`data/styles.json`](../data/styles.json) and at the top of each card; this page keeps only what holds across styles. Method and limits: [research-method.md](research-method.md).
 
+Added on 2026-10-10: 5 kinetic typography and lyric videos (style `tipografia-cinetica`, all 16:9, a 9th niche group: music). They are in the per-style numbers and their card, but not in the cross-style aggregates below, which still describe the 99 vertical shorts.
+
 ## Sound effects are rare
 
 | | Cuts with a sound effect on them |

package/src/skills-lib/wizz-video-styles/references/style-picker.md

@@ -7,6 +7,7 @@ The material decides which styles are **possible**. Niche and goal only rank the
 | What the project has | Possible styles |
 |---|---|
 | Voice only (own recording or TTS), no one on camera | [motion-editorial](styles/motion-editorial.md), [explicador-animado](styles/explicador-animado.md), [meme-rapido](styles/meme-rapido.md) |
+| A song with its lyrics, or a voice-over or slogan with no footage at all | [tipografia-cinetica](styles/tipografia-cinetica.md) (the words are the image; scenes generated in code) |
 | Voice + app/site screens | [motion-editorial](styles/motion-editorial.md) (screenshots in a phone card), [camera-no-setup](styles/camera-no-setup.md) (film the screen), [meme-rapido](styles/meme-rapido.md) |
 | A person on camera talking, nothing else | [talking-head-legenda](styles/talking-head-legenda.md) |
 | A person on camera + headlines, articles, charts, b-roll | [talking-head-cards](styles/talking-head-cards.md) (face bubble over full-screen b-roll is a variant) |
@@ -31,18 +32,19 @@ From [`data/references.json`](../data/references.json): the styles measured in e
 | dev / AI / build in public (EN) | meme-rapido (8.8M), camera-no-setup (8.1M), talking-head-legenda (5.4M), quadro-branco (4.4M), motion-editorial (0.83M), talking-head-cards (0.82M), podcast-clip (0.48M) |
 | dev / AI / tech business (BR) | quadro-branco (1.1M), talking-head-cards (0.41M), talking-head-legenda (0.26M), meme-rapido (0.23M), camera-no-setup (0.11M) |
 | personal finance / investing | podcast-clip (1.2M), talking-head-cards (1.2M), talking-head-legenda (0.77M) |
-| product / SaaS / marketing / business / law | desafio-numero (20.9M), talking-head-legenda (9.8M), camera-no-setup (0.53M), podcast-clip (0.33M), quadro-branco (0.15M), talking-head-cards (0.09M) |
+| product / SaaS / marketing / business / law | desafio-numero (20.9M), talking-head-legenda (9.8M), tipografia-cinetica (1.2M), camera-no-setup (0.53M), podcast-clip (0.33M), quadro-branco (0.15M), talking-head-cards (0.09M) |
 | health / fitness / nutrition | explicador-animado (33.4M), podcast-clip (9.8M), quadro-branco (0.61M), macro-asmr (0.54M), talking-head-cards (0.43M), desafio-numero (0.22M) |
 | food / recipes / lifestyle / travel | desafio-numero (107M), pov-vlog (8.7M), macro-asmr (7.7M), tour (1.05M) |
 | beauty / fashion / local services (clinic, salon, architecture, real estate) | macro-asmr (10.6M), talking-head-legenda (3.8M), talking-head-cards (2.0M), pov-vlog (1.4M), tour (0.37M) |
-| education / science / comedy | experimento (49M), explicador-animado (4.7M), esquete (2.9M), talking-head-cards (2.2M), talking-head-legenda (1.7M) |
+| education / science / comedy | experimento (49M), explicador-animado (4.7M), esquete (2.9M), talking-head-cards (2.2M), talking-head-legenda (1.7M), tipografia-cinetica (0.93M) |
+| music / artists / lyric videos | tipografia-cinetica (0.24M) |
 
 BR local services (clinic, salon, law, architecture), BR travel and BR sales/marketing were filled with a second pass (300k to 10.6M views). Still thin: BR software/SaaS proper (only sales/marketing found), real-estate agents in BR, and every niche has 9 to 16 references, not hundreds. When a niche is missing, pick by material and by the closest niche, and say the reference is borrowed.
 
 ## 3. Goal → tie-breaker
 
 - **Saves and comments (teach, explain a system):** motion-editorial, explicador-animado, quadro-branco, talking-head-cards.
-- **Reach and shares (curiosity, humour):** desafio-numero, experimento, meme-rapido, esquete, pov-vlog.
+- **Reach and shares (curiosity, humour):** desafio-numero, experimento, meme-rapido, esquete, pov-vlog, tipografia-cinetica (music and manifestos).
 - **Trust and authority (a person the audience should know):** talking-head-legenda, podcast-clip, camera-no-setup.
 - **Desire for a product:** macro-asmr, desafio-numero, camera-no-setup.
 

package/src/skills-lib/wizz-video-styles/references/styles/tipografia-cinetica.md

@@ -0,0 +1,68 @@
+# Kinetic typography and lyric video (mexicat P(doom), PixelValet, Ninh Truong)
+
+Measured on 5 shorts (en) in: product / SaaS / marketing / business / law, education / science / comedy, music / artists / lyric videos.
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 1,215,529 · median 353,630 |
+| Cuts per 10 s | min 0.63 · median 3.57 · max 5.54 |
+| Median scene length | 1.54 s |
+| Music bed (continuous / faint / none) | 3 / 0 / 2 |
+| Cuts with a sound effect on them | 31/94 (33%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+An audio track with words: a song with its lyrics, or a voice-over with its script. Nobody on camera, no footage. Everything is generated (HyperFrames, three.js, canvas). Word timing from ctc-align; a beat grid when there is music.
+
+## When it fits
+
+Music releases and lyric videos, manifestos, a strong opinion read aloud, brand promos built on a slogan, poetry. Shares and rewatches; weak for teaching step by step.
+
+## Recipe
+
+**Look**
+
+- The words ARE the image: one keyword huge in the frame, the rest of the line small. No caption strip at the bottom.
+- Plain background (black, or a flat colour that changes per section) or one generated world the camera travels through (a map, a grid in perspective, a 3D object the line wraps around).
+- Each word is animated by what it means: PAIN breaks, CHEERFUL bounces, LIGHT lights up, a word curves around the object it names.
+- Code-rendered pieces add film treatment: bloom, halation, grain, a thin HUD in mono.
+
+**Palette:** Dark base (#0A0A0A or a deep teal #0F3B3A) plus ONE glowing accent (orange #FF6A1A in P(doom), coral in Maps) and off-white text. Flat-colour variant: one saturated background per word or section, text in white or ink.
+
+**Typography:** Two or three families with high contrast: a heavy grotesk or condensed caps for the keyword (Archivo Black, Bebas Neue), a mono for small lines and HUD (IBM Plex Mono), optionally a serif italic for a soft word (Cormorant Garamond). Mix weights inside the same line.
+
+**Motion**
+
+- Words enter on the syllable they are sung or spoken (forced alignment, never guessed); the keyword lands on the beat.
+- Scene changes snap to the downbeat. P(doom) changes every ~0.3 s at peaks; the calm variant is a single continuous camera move for a whole verse.
+- The camera moves through the type (push, orbit, travel across a map) instead of the text sliding over a still frame.
+- Hold the last word of a line a beat longer before it leaves.
+
+**Captions:** None separate: the type is the caption. Keep the full line readable somewhere (small, mono, corner) when the keyword alone would lose the sense.
+
+**Hook (0 to 3 s):** The first sung or spoken word, huge, in the accent colour, landing on the first downbeat (0 to 1 s). No logo intro.
+
+**CTA:** The title or slogan locked up big and still for 1 to 2 s, the accent lit; one small line (artist, handle, link) under it.
+
+**Music:** The track itself is the whole soundtrack. Lyric videos: song at full level, no extra bed. Voice-over promos: a beat-driven bed under the voice.
+
+**SFX:** Song-driven pieces: almost none (5 of 56 cuts in P(doom), 0 of 2 in Maps). Promo variant: hits on every word change (9/13 and 16/20 cuts), which the heuristic partly confuses with the kick. Never add whooshes over a song.
+
+## Do not
+
+- Karaoke subtitles at the bottom over a static background: that is a lyric 'overlay', not this style.
+- Guessing word timing by ear; align it (ctc-align) and snap scenes to the beat grid.
+- More than one accent colour, or a different font for every word.
+- Shipping 16:9: every measured reference is horizontal; recompose for 9:16 with the keyword in the upper-middle and nothing in the bottom 20%.
+
+## Watch before editing
+
+- [PiXeL ARTWORKS](https://www.youtube.com/watch?v=9ngNySFSWWU) · 1,215,529 views
+- [PixelValet](https://www.youtube.com/watch?v=Kt8vGm8_ekQ) · 928,526 views
+- [Cinthya Studios](https://www.youtube.com/watch?v=A0CKFb24MJ8) · 353,630 views
+- [Ninh Truong](https://www.youtube.com/watch?v=FOMWS5pChuA) · 243,944 views
+- [mexicat](https://www.youtube.com/watch?v=5EoO5413dBY) · 120,706 views

package/src/skills-lib/wizz-video-styles/scripts/analyze.sh

@@ -9,7 +9,8 @@ id="$1"; out="$2"
 [[ "$id" =~ ^[A-Za-z0-9_-]{11}$ ]] || { echo "id inválido: $id" >&2; exit 2; }
 mkdir -p "$out"; cd "$out"
 f="./$id"  # prefixo ./ : ID que começa com hífen não vira opção do ffmpeg
-trap 'rm -f "$f.wav" "$f.mp4"' EXIT
+# .part: a first download that fails (HTTP 403) leaves it behind before the android fallback runs.
+trap 'rm -f "$f.wav" "$f.mp4" "$f".*.part' EXIT
 # Python with librosa + numpy: set VIDEO_STYLES_PY to a venv python, or python3 must have them.
 PY="${VIDEO_STYLES_PY:-python3}"
 "$PY" -c "import librosa, numpy" 2>/dev/null || { echo "precisa de librosa e numpy: pip install librosa numpy (ou VIDEO_STYLES_PY=/caminho/do/python)" >&2; exit 2; }

package/src/skills-lib/wizz-video-styles/scripts/beat_grid.py

@@ -0,0 +1,68 @@
+"""Beat grid of a song: tempo, beats and downbeats, so scene cuts can snap to the music.
+
+Usage: python3 beat_grid.py <audio_file> [--meter 4] [--fps 30] > grid.json
+
+Prints JSON: {"bpm", "meter", "duration_s", "beats": [s...], "downbeats": [s...],
+"downbeat_frames": [frame...]}. Needs librosa + numpy (same Python as analyze.sh;
+set VIDEO_STYLES_PY and call it with that interpreter).
+
+Downbeats are estimated, not detected: of the `meter` possible phases, the one
+whose beats carry the most low-frequency onset energy is taken as beat 1. Check
+two or three of them by ear before cutting a whole film on the grid.
+"""
+
+import argparse
+import json
+import sys
+
+import librosa
+import numpy as np
+
+HOP = 512
+LOW_HZ = 150
+
+
+def downbeat_phase(y_perc, sr, beat_frames, meter):
+    spec = np.abs(librosa.stft(y_perc, hop_length=HOP))
+    freqs = librosa.fft_frequencies(sr=sr)
+    low_energy = spec[freqs < LOW_HZ].sum(axis=0)
+    at_beats = low_energy[np.minimum(beat_frames, len(low_energy) - 1)]
+    scores = [at_beats[phase::meter].mean() if len(at_beats[phase::meter]) else 0.0 for phase in range(meter)]
+    return int(np.argmax(scores))
+
+
+def beat_grid(path, meter, fps):
+    y, sr = librosa.load(path, sr=22050, mono=True)
+    _, y_perc = librosa.effects.hpss(y)
+    _, beat_frames = librosa.beat.beat_track(y=y_perc, sr=sr, hop_length=HOP)
+    if len(beat_frames) < meter:
+        sys.exit(f"too few beats found in {path}: {len(beat_frames)} (is there a steady rhythm?)")
+    beats = librosa.frames_to_time(beat_frames, sr=sr, hop_length=HOP)
+    phase = downbeat_phase(y_perc, sr, beat_frames, meter)
+    downbeats = beats[phase::meter]
+    return {
+        # librosa's tempo and single beat intervals are quantised to the hop; the slope of a
+        # line fitted through every beat time averages that out.
+        "bpm": round(60.0 / float(np.polyfit(np.arange(len(beats)), beats, 1)[0]), 3),
+        "meter": meter,
+        "duration_s": round(len(y) / sr, 3),
+        "beats": [round(float(t), 3) for t in beats],
+        "downbeats": [round(float(t), 3) for t in downbeats],
+        "downbeat_frames": [int(round(t * fps)) for t in downbeats],
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    parser.add_argument("audio")
+    parser.add_argument("--meter", type=int, default=4)
+    parser.add_argument("--fps", type=int, default=30)
+    args = parser.parse_args()
+    if args.meter < 2 or args.fps < 1:
+        sys.exit("--meter must be >= 2 and --fps >= 1")
+    json.dump(beat_grid(args.audio, args.meter, args.fps), sys.stdout, indent=1)
+    print()
+
+
+if __name__ == "__main__":
+    main()